Command line

Verb	Does
packset ensure	start the writer if it is down, print PACKSET_URL
packset start / stop	the writer alone
packset status [WS]	up or down; counts by kind, the pin, the index and dense state
packset port / url / which	the port, the URL, the daemon binary
packset remember [--workspace WS] TEXT	one lesson
packset prefer [--workspace WS] TEXT	one standing preference
packset search [--workspace WS] QUERY	ranked claims: score, kind, id, text
packset due [WS]	claims whose review clock has run out, soonest first
packset islands [WS]	the link graph’s clusters, largest first: size and first claim
packset hubs [WS]	the claims the link graph turns on, highest first: a weighted PageRank over the links
packset island [--workspace WS] [--fire] CUE	the memories a cue activates: activation, seed mark, id, text; --fire wires the top eight together
packset fire [--workspace WS] ID ID...	these claims fired together; their links gain weight, their other links lose a little
packset grade ID [--lapsed] [WS]	mark a review recalled, or lapsed
packset pin [NAME]	read or set the pinned set; an empty name clears it
packset accessions [WS]	deed accessions the live claims cite
packset atoms [--as-of TS] [WS]	live-now claims, or those live at TS
packset citers ACCESSION [WS]	the live claims citing one accession
packset export --into DIR [WS]	claims to DIR/<ws>.jsonl; cited accessions to stdout

A workspace argument defaults to PACKSET_WORKSPACE, else the git remote of the working directory, else default.

HTTP

The writer listens on 127.0.0.1 only. Bodies are JSON.

Method and path	Does
GET /health	packsetd
GET /v1/status?workspace=	counts, pin, index and dense state
GET /v1/workspaces	names with live counts
POST /v1/atoms	store one claim; returns the stored record, or the live duplicate
GET /v1/atoms?workspace=&as_of=	live claims, or those live at as_of
GET /v1/atoms/<id>?workspace=	one record, whatever its state
POST /v1/atoms/update	merge fields into one live claim
POST /v1/atoms/delete	tombstone; why names the deed that withdrew it
POST /v1/grade	recalled true or false; moves due_at
GET /v1/search?workspace=&q=&limit=&as_of=&rerank=	ranked hits, the engine, the stage
GET /v1/islands?workspace=	the link graph’s communities, largest first
GET /v1/hubs?workspace=&limit=	the claims the link graph turns on, highest first, with their scores
GET /v1/activate?workspace=&q=&limit=&fire=	the cluster a cue activates: top hits as seeds, two hops along the links; fire=1 wires the top eight
POST /v1/fire	ids that fired together: weights up, missing links made, other links decayed
POST /v1/consolidate	the replacement rule over the live set in written order; apply false reports the pairs, true closes the earlier of each
GET /v1/accessions?workspace=	accessions the live claims cite
GET /v1/citers?workspace=&accession=	claims citing one accession
GET/POST /v1/pin	the pinned set
GET/POST /v1/set	a set’s three cards
GET /v1/pack?workspace=	the cards and the atoms as one document
GET/POST /v1/user, /v1/memory	the seat card, the workspace card
POST /v1/attach	hold one body for the next turn
GET/POST /v1/proposals, POST /v1/proposals/accept, POST /v1/compact	proposals a person accepts into claims

A refusal is a 4xx with error as one line: the reason a claim was not stored, quoted back.

The atom

Schema inside.atom/v1. Fields the writer reads:

Field	Meaning
kind	one of voice, habit, cache-pointer, preference, lesson, goal, conclusion, card_line, summary, correction, belief, trust, persona
level	explicit for a claim a person or seat wrote
text	the claim: at most two sentences of at most twenty-five words
workspace	required
set	optional; a checked name
entities	names the claim is about; an entity shaped like an accession must be one
ts, valid_from, valid_to	written, and the window it was true; valid_to closes on supersession
supersedes	the ids a contrary claim closed: an explicit id, a correction sharing an entity, a rewrite (token Jaccard 0.6), or the same opening words with a new object
due_at, review	the review clock: stability, difficulty, reps, last, interval_s
links	peers by entity overlap, at most eight, chosen by relative-neighbourhood pruning; fire adds peers up to eight
link_weights	weight per link in (0, 1], absent meaning 0.5; moved by fire
from, to, weight	a trust row: one edge of an influence graph, weight in (0, 1]; entities scope it to domains
name, anchor	a persona: a voter and how far it moves off its own ballot, in [0, 1]; the text is its view, entities its domains

Fields the writer does not model round-trip untouched.

The review clock

A new claim is due after one day. grade with recalled grows stability by how overdue the claim was; lapsed halves it and raises difficulty. The retrievability at t days since the last review with stability S is R = (1 + 19/81 * t/S)^(-1/2), so R(S) = 0.9. PACKSET_DECAY=fsrs scales a search score by R, floored at 0.25; cards are never scaled.

Environment

Variable	Read by	Meaning
PACKSET_URL	clients	the writer; INSIDE_MEMORY_URL is an alias
PACKSET_PORT	packset, packsetd	the port; default 8761
PACKSET_HOME	packsetd	the pack home; default ~/.grokinside/memory; two writers on one host want two homes and two ports
PACKSET_WORKSPACE	clients	the workspace when none is given
PACKSET_FUSE, PACKSET_DIVERSIFY, PACKSET_DECAY	writer	the panel
PACKSET_EMBED_QUERY_WORKERS	writer	query encoders kept side by side, default 2; each holds one model in memory
PACKSET_STEM	writer	off turns lexical stemming off
PACKSET_EMBED	writer	the encoder binary; else beside the writer, else on PATH
PACKSET_EMBED_CACHE	encoder	where models live; default $XDG_CACHE_HOME/packset/embed
PACKSET_RERANK, PACKSET_RERANK_MODEL	writer, encoder	the second stage on every search, and its model

Crates

Crate	Carries
packset-core	the atom, the prose limit, BM25+, the scan, the panel, decay, review clock
packset-daemon	packsetd: the Lightning Memory-Mapped Database (LMDB) store, HTTP, projections, the encoder child
packset-client	the HTTP client every tool uses
packset-cli	packset
packset-mcp	the read-only Model Context Protocol (MCP) surface
packset-embed	dense, late, sparse and cross-encoder models as one child process